ParaTempo cuts parallel reasoning latency by 21.8-32.2%

ParaTempo, a training-free framework introduced in a new arXiv preprint, reports a 21.8-32.2% cut in average latency and an 18.1-30.3% cut in total token usage for parallel-reasoning inference, with what the authors describe as 'competitive accuracy' on 'challenging mathematical and scientific reasoning benchmarks.'

The setup targets a familiar squeeze for teams running best-of-N or tree-style reasoning: cost 'grows with reasoning depth and branch count,' the paper notes. Existing schedulers rely on final-answer consensus, local token confidence, or isolated intermediate probes. Those signals, the authors argue, are 'often delayed, weakly tied to actual reasoning progress, or too noisy for dynamic, branch-level control.'

ParaTempo replaces them with a single score it calls 'temporal confidence,' defined as 'a branch-local measure of answer-space convergence,' or how sharply recent intermediate probes cluster on a dominant answer over time. Once evidence has accumulated, that one score drives the whole control loop: low-confidence branches are pruned, branches that persistently commit to their dominant answer are retired early, freed computation is reallocated by forking new branches, and generation halts globally once 'the confidence-weighted vote concentrates.'

The abstract names 'challenging mathematical and scientific reasoning benchmarks' but does not list them, does not name the base models tested, and reports accuracy only as 'competitive' rather than with numbers. The authors close on the claim that temporal confidence 'exhibits stronger temporal stability and predictive power for future branch convergence than token-level and instantaneous signals.'